Varsou Dimitra-Danai, Banerjee Arkaprava, Roy Joyita, Roy Kunal, Savvas Giannis, Sarimveis Haralambos, Wyrzykowska Ewelina, Balicki Mateusz, Puzyn Tomasz, Melagraki Georgia, Lynch Iseult, Afantitis Antreas
NovaMechanics MIKE, Piraeus 18545, Greece.
Entelos Institute, Larnaca 6059, Cyprus.
Beilstein J Nanotechnol. 2024 Nov 29;15:1536-1553. doi: 10.3762/bjnano.15.121. eCollection 2024.
A key step in building regulatory acceptance of alternative or non-animal test methods has long been the use of interlaboratory comparisons or round-robins (RRs), in which a common test material and standard operating procedure is provided to all participants, who measure the specific endpoint and return their data for statistical comparison to demonstrate the reproducibility of the method. While there is currently no standard approach for the comparison of modelling approaches, consensus modelling is emerging as a "modelling equivalent" of a RR. We demonstrate here a novel approach to evaluate the performance of different models for the same endpoint (nanomaterials' zeta potential) trained using a common dataset, through generation of a consensus model, leading to increased confidence in the model predictions and underlying models. Using a publicly available dataset, four research groups (NovaMechanics Ltd. (NovaM)-Cyprus, National Technical University of Athens (NTUA)-Greece, QSAR Lab Ltd.-Poland, and DTC Lab-India) built five distinct machine learning (ML) models for the in silico prediction of the zeta potential of metal and metal oxide-nanomaterials (NMs) in aqueous media. The individual models were integrated into a consensus modelling scheme, enhancing their predictive accuracy and reducing their biases. The consensus models outperform the individual models, resulting in more reliable predictions. We propose this approach as a valuable method for increasing the validity of nanoinformatics models and driving regulatory acceptance of in silico new approach methodologies for the use within an "Integrated Approach to Testing and Assessment" (IATA) for risk assessment of NMs.
长期以来，在监管部门接受替代或非动物试验方法的过程中，一个关键步骤是进行实验室间比对或循环试验（RRs），即向所有参与者提供一种通用的测试材料和标准操作程序，参与者测量特定终点并返回数据进行统计比较，以证明该方法的可重复性。虽然目前尚无比较建模方法的标准方法，但共识建模正成为RR的“建模等效物”。我们在此展示一种新颖的方法，通过生成共识模型来评估使用通用数据集训练的针对同一终点（纳米材料的zeta电位）的不同模型的性能，从而增强对模型预测和基础模型的信心。利用一个公开可用的数据集，四个研究小组（塞浦路斯的NovaMechanics有限公司（NovaM）、希腊的雅典国立技术大学（NTUA）、波兰的QSAR实验室有限公司和印度的DTC实验室）构建了五个不同的机器学习（ML）模型，用于在水介质中对金属和金属氧化物纳米材料（NMs）的zeta电位进行计算机模拟预测。将各个模型整合到一个共识建模方案中，提高了它们的预测准确性并减少了偏差。共识模型的表现优于各个模型，从而得出更可靠的预测结果。我们建议将这种方法作为一种有价值的方法，以提高纳米信息学模型的有效性，并推动监管部门接受在“综合测试与评估方法”（IATA）中用于纳米材料风险评估的计算机模拟新方法。
